is f_jit(x) different compared to jax.jit(f)(x)

[rozeappletree]

selu_jit = jax.jit(selu)
%timeit -n1000 -r5 selu_jit(x).block_until_ready()

1000 loops, best of 5: 122 µs per loop

and

%timeit -n1000 -r5 jax.jit(selu)(x).block_until_ready()

1000 loops, best of 5: 270 µs per loop

Is the difference significant

[jakevdp]

The difference here is the time that it takes the jax.jit transform to do a cache lookup on selu and construct & return a wrapped function. You can measure this more directly if you wish:

%timeit jax.jit(selu)

As for whether it's significant – it's a one-time cost of several tens of microseconds. When running microbenchmarks, this can be significant. If you're JIT-compiling a larger/more complex piece of code, it may not be significant compared to the cost of the full computation.